akniele/classify_traffic_signs

classify_nl_traffic_signs

YOLOv26 image classification model for identifying Dutch traffic sign classes. Trained on cropped sign images from the Panoramax platform.

This model is the second stage in a two-step pipeline:

1. Detection — Panoramax/detect_face_plate_sign localises the sub-sign crop
2. Classification — this model identifies the traffic sign class from the crop

Which signs are included?

The model includes the most common traffic signs in the Netherlands, such as warning signs, priority signs, parking signs, and speed signs. It classifies the signs by the Dutch RVV-code (RVV 1990), for example A1 for a speed limit sign. A full overview of these classes can be found below. In addition, the model also includes some signs that are not strictly traffic signs but are nonetheless detected as a sign by the detection model. Some of these include dry standpipe connection points (droge blusleidingen), chevron signs (bochtschilden), channelizing panels (geleidebakens), a few nautical traffic signs such as max speed for boats, and no-alcohol signs.

Model details

Property	Value
Architecture	YOLO26s-cls (Ultralytics)
Base model	yolo26s-cls.pt (pretrained ImageNet)
Fine-tuning epochs	30
Input size	224 × 224 px
Classes	170
Training images	19,374
Validation images	4,760

Classes

The following are the classes that have an official code:

ID	Class	Description
0	A04-15	adviessnelheid_15
1	A04-30	adviessnelheid_30
2	D02-LO	arrow_left
3	C04-L	arrow_left_rect
4	D03	arrow_left_right
5	D02-RO	arrow_right
6	C04-R	arrow_right_rect
7	D04	arrow_through
8	D05-L	arrow_turn_left
9	D07	arrow_turn_left_right
10	D06-L	arrow_turn_left_through
11	D05-R	arrow_turn_right
12	D06-R	arrow_turn_right_through
13	G12	bicycle_end
14	L1002-A	bicycle_priority
15	G11	bicycle_start
16	G12-B	moped_end
17	G12-A	moped_start
18	BW410	bike_route
19	BB12	bochtschild
20	C05	bothways
21	BW03	bromfiets_hier
22	F13	bus_start
23	F14	bus_end
24	L05	doorgaand_plus_einde_rijstrook
25	L06	doorgaand_plus_splitsing
26	G05	erf
27	BB01	exit
28	BB35	geleidebakens
29	B06	give_way
30	J37	hazard
31	J30	hazard_airport
32	J28	hazard_animals
33	J24	hazard_bicycle
34	J39	hazard_bollard
35	J29	hazard_bothways
36	J38	hazard_bump
37	J01	hazard_bumps
38	J08	hazard_crossing
39	J26	hazard_dock
40	J25	hazard_gravel
41	J36	hazard_ice
42	J07	hazard_incline
43	B03	hazard_intersection
44	B04	hazard_intersection_left
45	B05	hazard_intersection_right
46	J03	hazard_left_turn
47	J10	hazard_level_crossing
48	C1-snor	hazard_moped
49	J15	hazard_moving_bridge
50	J17	hazard_narrow_both
51	J19	hazard_narrow_left
52	J18	hazard_narrow_right
53	J22	hazard_pedestrian_crossing
54	J23	hazard_pedestrian
55	J16	hazard_road_works
56	J1004	hazard_rocks
57	J09	hazard_roundabout
58	J21	hazard_school
59	J20	hazard_slippery
60	J32	hazard_traffic_signal
61	J11	hazard_train
62	J14	hazard_tram
63	J02	hazard_right_turn
64	J27	hazard_wild_animals
65	J04	hazard_zigzag
66	J05	hazard_zigzag_left
67	G09	horse_start
68	C20	maxaxleweight
69	C19	maxheight
70	C17	maxlength
71	A01-5	maxspeed_5
72	A01-10	maxspeed_10
73	A01-15	maxspeed_15
74	A01-20	maxspeed_20
75	A01-25	maxspeed_25
76	A02-30	maxspeed_30_end
77	A01-40	maxspeed_40
78	A01-50	maxspeed_50
79	A02-50	maxspeed_50_end
80	A01-60	maxspeed_60
81	A02-60	maxspeed_60_end
82	A01-70	maxspeed_70
83	A02-70	maxspeed_70_end
84	A01-80	maxspeed_80
85	A02-80	maxspeed_80_end
86	A01-90	maxspeed_90
87	A01-100	maxspeed_100
88	A02-100	maxspeed_100_end
89	A01-120	maxspeed_120
90	A01-30	maxspeed_30
91	C21	maxweight
92	C18	maxwidth
93	C22-E	milieu
94	G02	motorway_end
95	G01	motorway_start
96	C08	no_agriculture
97	C09	no_agriculture_bicycle
98	C22-C	no_bestelbus_vrachtwagen
99	C14	no_bicycle
100	C07-A	no_bus
101	C12	no_car_motorcycle
102	C06	no_car
103	C22	no_dangerous
104	C02	no_entry
105	L08	noexit
106	C07	no_hgv
107	C13	no_moped
108	C11	no_motorcycle
109	F01	no_overtaking
110	F02	no_overtaking_end
111	F03	no_overtaking_hgv
112	F04	no_overtaking_hgv_end
113	E03	no_parking_bicycle
114	E01	no_parking
115	C16	no_pedestrian
116	E02	no_stop
117	F07	no_u_turn
118	C01	no_way
119	L09-2L	noexit_left
120	F08	no_end
121	T201	omleiding
122	C03	oneway
123	G13	onverplicht_fietspad
124	E09	parkeren_vergunninghouders
125	E04	parking
126	E08-9	parking_bicycle
127	E08	parking_car
128	E102-2	parking_card
129	E08o	parking_charging
130	E105	parking_covered
131	E07	parking_delivery
132	E06	parking_disabled
133	E08-7	parking_moped
134	E101-2	parking_payment
135	E08-11	parking_police
136	E05	parking_taxi
137	G08	pedestrian_end
138	G07	pedestrian_start
139	L02	pedestrian_crossing
140	B01	pedestrian_start
141	F05	priority_backward
142	B02	priority_end
143	F06	priority_forward
144	L14	rest
145	D01	roundabout
146	B07	stop
147	F15	tram_start
148	F17	tram_bus
149	G04	trunk_end
150	G03	trunk_start
151	BB04	versmalling_vluchtstrook
152	L04	voorsorteren

As mentioned before, there are also a number of other classes that are not strictly traffic signs. Further, there are a number of bad classes such as windows, buildings and the backs of traffic signs. These accidentally get detected as signs in step 1 of the pipeline.

Design choices: numbers on signs and zone signs

As stated above, the model classifies signs by RVV-code, such that it can distinguish between for instance a zebra-crossing sign and a weight limit sign. However, there are no separate classes for different numerical values on the signs, for instance a weight limit sign with weight limit 7.5 tons vs. a weight limit sign with weight limit 20 tons. There is a separate model for reading the numerical values on signs, found at Panoramax/????. Speed limit signs are an exception: For these, separate classes are included per speed limit.

There are also no separate classes for whether a sign is a zone sign or not. For instance, a speed limit 30 sign in the Netherlands can be a regular sign, or a sign indicating the start, end, or repetition of a speed limit zone. Both of these signs will receive the class A1 in this model. To distinguish between them, there is a separate model that classifies signs by whether they are a zone-start, zone-repetition, zone-end, or no-zone sign. The model can be found at Panoramax/????.

Performance

Validation set (4,760 images)

Metric	Value
Top-1 accuracy	98.4%
Top-5 accuracy	99.8%

Usage

from ultralytics import YOLO

model = YOLO("Panoramax/????")

results = model.predict("sign_crop.jpg", conf=0.50)
for r in results:
    print(r.probs.top1, model.names[r.probs.top1], float(r.probs.top1conf))

The model expects a cropped image of a sign — not a full street photo. Use Panoramax/detect_face_plate_sign to obtain the crop first.

Predictions of classes ????? should be forwarded to XXXX. Predictions of classes ??? should be forwarded to XXXX.

Training

Parameter	Value
Epochs	30
Batch size	16
Image size	224 × 224
Optimizer	AdamW
Initial LR	0.001
Augmentation	RandAugment, fliplr disabled, translate=0.2, shear=10.0, erasing=0.2, dropout=0.3

The training dataset used for this model is based on the dataset Panoramax/Panoramax/classified_nl_road_signs. The dataset was extended with additional Dutch traffic sign and non-traffic sign classes. Zone- and non-zone signs of the same RVV code were merged into a single class.

Dataset

See Panoramax/??? for the training and validation images.

License

Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)